Over the past few months, OpenAI has made a massive purchase — tens of thousands of Mac mini and Mac Studio computers. This is not just an upgrade of its hardware fleet, but a strategic move aimed at solving specific tasks related to training AI agents. We are talking about so-called computer-use agents — systems that learn to independently interact with desktop operating systems: clicking, typing, and managing the interface. It is precisely for these purposes that reinforcement learning is used, which requires a huge number of parallel environments.

Notably, my main competitor in this race, Anthropic, is solving a similar problem differently. The company is not buying hardware but renting Mac mini capacity from Amazon Web Services. This highlights different approaches to scaling: some prefer control over hardware, while others prefer the flexibility of cloud solutions.

Interestingly, on August 25, literally a few days before this information appeared in the public domain, Apple unveiled an updated lineup of devices based on the M5 Ultra and M6 chips. A coincidence? Hardly. An analysis of the stated technical specifications clearly points to a division of roles for these computers in the AI laboratories of the future.

Why Mac, not GPU clusters?

  • Mac mini M6. The compact form factor is an ideal solution for deploying thousands of independent virtual environments. AI agents need real operating systems for training, not emulation. This allows them to practice skills in conditions as close to reality as possible.
  • Mac Studio M5 Ultra. Support for up to 512 GB of unified memory is a critical reserve for locally running and hosting the heaviest language models. For example, the flagship model with a bandwidth of 1.2 TB/s can entirely fit the open-source LLM GLM-5.3-Flash with 320 billion parameters, which, depending on quantization, can take up to 300 GB of RAM.

For comparison: on a regular desktop, such an amount of video memory would require installing a setup of 10 flagship RTX 5090 graphics cards with 32 GB each. This entails colossal electricity costs and a significant loss of bandwidth due to inter-processor connections. Apple solved this problem elegantly and efficiently.

It is noteworthy that back on April 30, 2026, Tim Cook, who then held the position of Apple's CEO, publicly acknowledged the shortage of Mac mini and Mac Studio, directly citing agentic AI tools as the reason. The company has already adapted its marketing, positioning the new devices as solutions for "continuously running local agentic computing."

It is important to understand: Apple hardware is not intended to replace the Nvidia GPU clusters on which base models are trained. Instead, it forms a new infrastructure layer — a virtual "office" where trained neural networks practice performing real tasks. Neither OpenAI nor Anthropic has disclosed the exact number of devices, the breakdown by model, or the purchase amounts, but the scale is obvious.

My analysis: We are witnessing the emergence of a new market — infrastructure for AI "practice." While all attention is focused on data centers for training, the battle for the agents' "habitat" has already begun. And Apple, perhaps unexpectedly, could become a key player in this segment, monetizing not only hardware but also an ecosystem that is becoming the de facto standard for agentic computing.